char-fork/llama.cpp
git clone https://git.t4t.associates/char-fork/llama.cpp
fcae601e4
char/aliceai-t5-moe
1#version 450 2 3#include "types.glsl" 4#include "generic_unary_head.glsl" 5 6layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; 7 8void main() { 9 const uint idx = get_idx(); 10 11 if (idx >= p.ne) { 12 return; 13 } 14 15#if defined(DATA_A_BF16) 16 data_d[get_doffset() + dst_idx(idx)] = D_TYPE(bf16_to_fp32(uint32_t(data_a[get_aoffset() + src0_idx(idx)]))); 17#elif defined(DATA_D_BF16) 18 float f = float(data_a[get_aoffset() + src0_idx(idx)]); 19 data_d[get_doffset() + dst_idx(idx)] = D_TYPE(fp32_to_bf16(f)); 20#elif !defined(OPTIMIZATION_ERROR_WORKAROUND) 21 data_d[get_doffset() + dst_idx(idx)] = D_TYPE(data_a[get_aoffset() + src0_idx(idx)]); 22#else 23 data_d[get_doffset() + dst_idx(idx)] = data_a[get_aoffset() + src0_idx(idx)]; 24#endif 25}